TL;DR
In 2025 about a third of companies told McKinsey that AI would reduce their headcount. A year later, one in seven said it actually had. Then the same survey found that four in ten expect cuts in the coming year. The forecast did not learn from its own error, because the forecast is not really a forecast. Meanwhile something real is happening, and it is narrower and quieter than the announcements: not people leaving, but people not arriving.
The number that keeps getting louder as it gets smaller
On 3 September 2026 McKinsey published a chart that almost nobody shared.
In 2025, 32 percent of the executives it surveyed expected AI to reduce their organisation's headcount. A year later, 14 percent reported that any reduction had actually happened. Two thirds reported little or no AI-related change in total employment at all.
Then the same respondents were asked about the next twelve months. 39 percent expect AI-driven cuts.
The first forecast missed by more than half. The second forecast went up.
A press release is not a measurement
Challenger, Gray & Christmas counts layoffs by the reason employers give for them. In its July 2026 report, out on 6 August, AI was the leading stated reason for the fifth month running: 10.970 of 33.429 cuts that month, 112.713 so far in the year.
Read the firm's own caveat next to its own number. AI-related cutting, it says, "has been limited outside of the Tech sector". And then this, which is the most useful sentence written about AI and work all year:
Naming AI in a layoff announcement can win over investors while pushing current and prospective employees away.
That is a company telling you the metric it sells is partly a market for stories. The tally does not measure what machines can do. It measures demand for a particular explanation, priced daily.
Oxford Economics put it more bluntly back in February: firms are "trying to dress up layoffs as a good news story rather than bad news". Martha Gimbel, who runs The Budget Lab at Yale, looked at the same period and said that no matter how you cut the data, there was no major macroeconomic effect to find.
Her tracker, updated on 15 September with August payroll data, still says the same thing. So does a survey of roughly 6.000 executives published by the NBER in February: 69 percent of their firms use AI, more than 90 percent report no employment impact, and the measured average effect on productivity is 0,29 percent. Not a typo. Zero point two nine.
The part that is actually true
Here is where the comfortable version of this story goes wrong, and I have no interest in the comfortable version.
Something is happening. It is just not a layoff.
The Census Bureau's Center for Economic Studies published a paper in September on graduates from AI-exposed majors: 5 percentage points lower initial employment rate, 13 percent lower earnings in the first full quarter. The authors compare the hit to graduating during a major recession. Stanford's Digital Economy Lab, using payroll records rather than surveys, finds employment for 22 to 25 year olds in highly exposed occupations running about 19 percent below the trend of their peers in unexposed work.
Both find the same mechanism: fewer hires, not more firings. The Dallas Fed, looking at the same question in January, calls it reduced inflow and adds the caveat everyone else drops, that the pattern may not be causal at all.
That is why the macro indicators stay flat. They were built to count people in jobs. Nobody built an indicator for a door that quietly stops opening.
Why this matters for your business
If you are planning around the announcements, you are planning around a genre.
The honest read of the 2026 data is this. Almost nobody is being replaced. Someone is not being hired, and it is the same someone everywhere: the person who would have learned the job by doing the easy parts of it, the parts that are now automated. That is a hiring pipeline problem with a five year fuse, not a cost line you can book this quarter.
Two things follow. Do not let a vendor price your project against headcount you were never going to cut. And look at who is missing from your org chart at the bottom, because in about five years you will need them in the middle.